Ken Pritchard created DBCP-601:
----------------------------------
Summary: LocalXAResource.prepare() leaves a read-only connection
enlisted after voting XA_RDONLY, so the next transaction fails to enlist
Key: DBCP-601
URL: https://issues.apache.org/jira/browse/DBCP-601
Project: Commons DBCP
Issue Type: Bug
Affects Versions: 2.14.0
Reporter: Ken Pritchard
When a transaction enlists a read-only {{BasicManagedDataSource}} connection
together with at least one other XA resource, the transaction manager runs
two-phase commit. {{LocalXAConnectionFactory.LocalXAResource.prepare()}}
returns {{XA_RDONLY}} for a read-only connection, but it does not clear
{{currentXid}}.
Under the XA specification, a resource that votes {{XA_RDONLY}} takes no part
in phase 2: the transaction manager does not call {{commit()}} or
{{rollback()}} on it. Those are the only places {{currentXid}} is cleared, so
the connection goes back to the pool still associated with a transaction that
has already completed.
The next transaction that borrows that connection fails in {{start()}}:
{noformat}
javax.transaction.xa.XAException: Already enlisted in another transaction with
xid ...
java.sql.SQLException: Unable to enlist connection in transaction:
enlistResource returns 'false'.
javax.transaction.RollbackException: Unable to commit: transaction marked for
rollback
{noformat}
The pool then destroys that connection, so the transaction after it succeeds.
Each read-only two-phase transaction therefore causes one later, unrelated
transaction to fail and roll back, which makes the cause hard to trace.
*Steps to reproduce*
# Configure a {{BasicManagedDataSource}} with {{defaultReadOnly=true}} and
{{maxTotal=1}}.
# Transaction 1: begin, get a connection, enlist a second {{XAResource}} that
votes {{XA_OK}}, commit. It commits normally through two-phase commit.
# Transaction 2: begin, get a connection. Enlisting fails as above, and commit
throws {{RollbackException}}.
# Transaction 3: works, on a new physical connection.
Reproduced on master (7b07108c and 4f68273a) with the Geronimo transaction
manager that DBCP's own tests use.
*Not affected:* single-resource transactions (one-phase commit calls
{{commit(xid, true)}}, which clears {{currentXid}}) and read-write connections.
The read-only branch in {{prepare()}} dates from DBCP-228. The existing tests
don't reach it because H2's {{Connection.isReadOnly()}} always returns
{{false}}.
The same mistake, clearing per-transaction state on commit and rollback but not
after a read-only prepare, has been fixed in other Apache projects: AMQ-2556
(ActiveMQ 5.4.0; transaction state "not released ... at prepare time where
result is XAResource.XA_RDONLY") and DERBY-5562 (Derby 10.8.3.0 / 10.9.1.0;
"When XA_RDONLY is returned, no commit or rollback will be received").
Apache Tomcat bundles this code as {{org.apache.tomcat.dbcp.dbcp2.managed}}, so
it has the same behavior until it next syncs from a DBCP release.
*Proposed fix:* on the read-only path in {{prepare()}}, clear {{currentXid}}
before returning {{XA_RDONLY}} (still returning {{XA_RDONLY}}, so the
transaction manager skips phase 2 as it should). I have a patch with tests and
will open a pull request.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)